Repository navigation
Releases
MeteorLLM v0.3.0 — Qwen3.5-0.8B + four-core speedup
MeteorLLM runs Qwen3.5-0.8B completely offline on PS Vita.
November 6 speedup update — four-core inference
Reinstall the updated MeteorLLM-v0.3.0.vpk to get this speedup. Existing 0.8B users do not need to download the model again.
The VPK now bundles GrapheneCt's MIT-licensed CapUnlocker v1.4 and uses the Vita's fourth CPU core for matrix and attention work. It loads the plugin at runtime when needed, falls back to three workers if unavailable, and does not edit taiHEN boot configuration or set CPU clocks. The plugin's system capability changes last until reboot. To force three workers and skip automatic loading, create ux0:data/VitaDesk/MeteorLLM-08/three-core.txt, reboot, and relaunch.
| Same short Tavi prompt | Through natural EOS |
|---|---|
| Three workers, matched controls | 34.45–34.54 s |
| Four workers, two repeats | 30.69–30.70 s |
That is about 11% less time, with the correct Tavi answer and identical full logits at every generated step. All four cores were observed doing work. The matched tests reported 444 MHz, not 500 MHz. Only the 32-token shared system/template prefix was cached; all 35 user/suffix prompt tokens were processed on every run. Startup is excluded. No question or answer was precomputed.
The installed normal UI also passed after reboot in 30.14 s, first token 25.84 s, with 4 MiB free user system RAM. Its effective clock was not independently logged, so use the matched controls for the speed comparison. DMR's larger cache saved only ~0.47 s and is not included. Model weights and quantization are unchanged.
Install or update
Install MeteorLLM-v0.3.0.vpk over the existing METEOR001 app. New users, or users upgrading from 0.6B, must also extract the four files from MeteorLLM-Qwen35-0.8B-model.zip directly into ux0:data/VitaDesk/MeteorLLM-08/. Old 0.6B files remain untouched in their old folder; reinstall v0.2.1 to roll back to that model. PSVshellPlus is not required.
The model and vocabulary shortlist remain approximate and can make factual, arithmetic, grammar, and instruction-following mistakes. This is a short performance/quality smoke test, not a broad stability or accuracy evaluation. No image-input support.
Source, evidence, and checksums
- Updated README
- Matched benchmark logs and methodology
- MeteorLLM-v0.3.0-four-core-source.zip matches source commit
7af5c0ecad70e0d8d4c5736edef498bc953d4ebe. The original v0.3.0 tag is preserved; GitHub's automatic “Source code” downloads predate this patch. - SHA256SUMS.txt contains the refreshed asset hashes.
- Tavi-four-core.png shows the updated app after its reboot test; the other screenshots show the earlier v0.3.0 build.
MeteorLLM v0.2.1
Vita runtime
The same 67-token Tavi benchmark finished in 37.14 seconds with the template ready, down from 39.00 seconds in v0.2.0. First token arrived in 28.18 seconds. Both runs returned Tavi. with token IDs 51, 6190, 13 and natural EOS.
Only the 32 shared chat-template tokens were cached. Preparing them took 20.80 seconds at startup. Template preparation plus the measured response took 57.94 seconds, excluding model load. The benchmark covers one prompt. See the README for the configuration, output tradeoffs, and overrides.
Install
Install the VPK over v0.2.0 and keep your existing model files. On a fresh setup, follow the README and get the compatible Q4 output archive from v0.2.0. The same output file enables the faster path in this release.
MeteorLLM v0.2.0
MeteorLLM v0.2.0 adds rolling conversations, configurable reply length, and faster local Qwen3-0.6B inference on PS Vita.
Vita runtime
The same 67-token formatted Tavi prompt returned Tavi. with token IDs 51, 6190, 13 and natural EOS.
| Build | First token | Submit to EOS |
|---|---|---|
| Earlier optimized build | 92.58 s | 136.38 s |
| Vocabulary shortlist, no template cache | 60.63 s | 69.57 s |
| Previous tiled default, template ready | 32.71 s | 41.69 s |
| v0.2.0, template ready | 30.07 s | 39.00 s |
The newest matrix kernel processes eight prompt tokens across SIMD lanes. The build also uses int8 activations, a Q4 output projection, and a vocabulary shortlist. Preparing the shared 32-token template took 22.72 s at startup. Preparation plus the measured response took 61.72 s, excluding model load. The question and generated answer were processed during the timed run.
This is a single short prompt on physical hardware. The shortlist and quantized math can change output on other prompts. See the README for the benchmark settings, clock measurement caveat, and full-vocabulary and cold-mode overrides.
Install
- Install
MeteorLLM.vpk. - Keep or prepare the original
weights.bin,embedding.bin, andtokenizer.binusing the repository instructions. - Extract
output-q4-test.binfromMeteorLLM-Q4-output.zipand put it inux0:data/VitaDesk/MeteorLLM/with the model files. This enables the faster output path for the tested model.
The output archive includes the upstream model license and a modification notice. The measured Vita run is attached as Tavi-Vita.log.
MeteorLLM v0.1.2 (touch scrolling)
Drag the front touchscreen over the conversation or hold Up or Down to scroll the full transcript. Verified both controls on physical Vita hardware with a long prompt and generated reply. Install the VPK over the previous version. Vita package version 00.03.
MeteorLLM v0.1.1 (chat UI fix)
This experimental update adds a scrollable transcript for the current chat. Hold Up or Down to read earlier prompts and full replies. While you read older text, incoming tokens leave your scroll position in place.
Install MeteorLLM.vpk over the previous app. The title ID remains METEOR001 and the Vita package version is 00.02. Follow the README for model setup. The 128-token reply cap and 512-token conversation window still apply.
Built with VitaSDK and checked that the packaged executable matches the build output. This UI patch has not been tested on a physical Vita.
MeteorLLM v0.1.0 (experimental)
Experimental PlayStation Vita build of MeteorLLM (title ID METEOR001).
Install MeteorLLM.vpk and follow the repository README to export and install the model files. The package contains the app. Model weights and tokenizer data are prepared separately.
For the measured setup, use the README's configuration files for three workers, batch eight, fused NEON, parallel attention, and 256-row output chunks. This build completed the 67-token ordering test in 136.381 seconds on a physical Vita at 333 MHz. Simple arithmetic can fail.